[人人能懂] AI的“开窍”瞬间，我们能学到什么？

00:00:34 让AI更“有谱儿”：不止一条路通罗马 

00:05:21 如何让AI更聪明？一个“求稳”的智慧 

00:09:14 造车新智慧：如何用“搬沙子”的办法，算出最省油的外形？ 

00:12:31 AI也会“路径依赖”？一个简单动作，让它“老树发新芽” 

00:17:01 AI炼丹术：我们如何“教会”机器遵守化学规则？ 

00:21:50 AI进化论：从“死记硬背”到“自我成长” 

本期介绍的几篇论文：

[LG] Flow Matching Policy Gradients  

[UC Berkeley]  

https://arxiv.org/abs/2507.21053 

---

[CL] Geometric-Mean Policy Optimization  

[Microsoft Research]  

https://arxiv.org/abs/2507.20673  

---

[LG] Geometric Operator Learning with Optimal Transport  

[California Institute of Technology & Nvidia]  

https://arxiv.org/abs/2507.20065  

---

[LG] What Can Grokking Teach Us About Learning Under Nonstationarity?  

[Google DeepMind]  

https://arxiv.org/abs/2507.20057  

---

[LG] Enhancing Materials Discovery with Valence Constrained Design in Generative Modeling  

[MIT]  

https://arxiv.org/abs/2507.19799  

---

[LG] A Survey of Self-Evolving Agents: On Path to Artificial Super Intelligence  

https://arxiv.org/abs/2507.21046